[2026-09-22] What Should a Self-Teacher See? Privileged Context Design for On-Policy Self-Distillation
这篇系统比较 self-teacher 的五档 privileged context,从 full solution 到 answer-only。本文逐项核对 29,434 行公开数据、L2-L4 compiler…
这篇系统比较 self-teacher 的五档 privileged context,从 full solution 到 answer-only。本文逐项核对 29,434 行公开数据、L2-L4 compiler…